Super User Daily: 2026-04-24
The day Anthropic quietly stripped Claude Code from the $20 Pro plan and got dragged into reversing it by noon was also the day super users showed what they actually do with these things. Below the pricing shouting match, people were rebuilding data recovery companies on a 16TB NAS, replicating an 18-year trading career on a Mac Mini in 48 hours, copying whale wallets into $14k monthly profit, and wiring Claude Code into everything from Figma design systems to BLE packet sniffers. The theme is consistent: the harness wins, the model is a commodity, and token budgets are the new salary line.
@marlowxbt [Claude Code]
https://x.com/marlowxbt/status/2047017269402726825
A 17-year-old in the US convinced his mom to buy a $280 used Steam Deck by telling her it runs Linux, never installed a single game, and turned it into his primary dev machine. He set up Claude Code with custom skills and hooks that auto-run tests on file save, then shipped four paid apps to local businesses in three weeks: a barber booking page, a vape shop inventory tracker, a taco truck menu, and a dog groomer scheduler. Month-one revenue: $13,000. He codes with his thumbs from bed while his friends play Fortnite on the same hardware.
@LunarResearcher [Claude Code]
https://x.com/LunarResearcher/status/2046971446325313876
An Anthropic agent team engineer recognized his Polymarket bot in a Sightglass cafe. The setup: Claude Code ingests 86 million trades across 14,000 wallets, ranks by win rate and profit, clones the top 47. Three agents (arbitrage, convergence, whale copy) share one wallet but no memory — 2 agree for full size, 1 alone for half, disagree skip. Consensus filter alone killed 40% of losing trades. $200 seed to $14,300 in 27 days, 74% win rate, Sharpe 2.47, zero touches since. Total monthly stack: $20 Claude API, $5 VPS, everything else free.
@pashmerepat [OpenClaw]
https://x.com/pashmerepat/status/2046865863979172039
Inside-baseball from an OpenClaw dev on a real harness regression. When OpenClaw was configured to use the Codex harness with OpenAI models, auth was silently broken and the system was falling back to the Pi harness without warning. Two PRs (fix auth bridge, kill silent fallback) flipped the agent from doing the same shallow loop every heartbeat — read file, check Discord, say "HEARTBEAT_OK" — to full loops that actually read workspace context, inspect the repo, make edits, verify, and pick up where they left off on restart. Same model, same prompts, just a working harness.
@yonann [OpenClaw]
https://x.com/yonann/status/2046896292341657997
Chris Camillo reports that at least six people used OpenClaw on a $650 Mac Mini to replicate his 18-year trading strategy in 48 hours via voice prompting. The agent scrapes TikTok videos and comments, scans for signal phrases like "sold out" and "obsessed," evaluates against his methodology, and trades 24/7. He watched his own playbook get rebuilt on consumer hardware by users who never met him. This is the OpenClaw pattern at its ugliest for anyone selling "my method" as a product: the method is now a folder of skills.
@om_patel5 [Claude Code]
https://x.com/om_patel5/status/2046753036463173994
After 20 years of losing files across multiple crashed drives, a user bought a 16TB NAS, pointed Claude Code at five corrupted drives, and had it reconstruct the original folder structures by inference — reading file contents to figure out what went where, then consolidating hundreds of thousands of loose files into a clean master library with automated backup. Data recovery companies charge $5,000+ for this and typically recover half. Claude Code did it in an hour on her own hardware.
@0xwhrrari [Claude Code]
https://x.com/0xwhrrari/status/2047006144565002382
A neighbor at Anthropic leaked a killed-internal-experiment: weather derivatives had the highest edge of any Polymarket category, 3x above anything else, shut down over alignment concerns. The user built it in one night with Claude Code: NOAA 6z model vs market odds, enter on spreads above 8%, exit when spread collapses under 3%. 312 markets traded, 74% win rate, $200 seed to $14,000 in 27 days, Sharpe 3.04. The edge is a 6-hour window between the 6z NOAA drop and the 12z market correction.
@mikefutia [Claude Code]
https://x.com/mikefutia/status/2047019993439850631
A static-ad system rebuilt inside Claude Code on the new ChatGPT Images 2.0 model: brand name + URL in, 40 production-ready static ads out. Claude researches brand fonts, colors, packaging, and photography style, builds a Brand DNA document, fills 40 proven ad templates (headline, us-vs-them, testimonial, UGC, review cards, stat callouts), fires prompts to Images 2.0 with product photos as reference, and downloads ads into organized folders with an HTML gallery. No Canva, no designer brief, no prompt filling.
@AlfieJCarter [Claude Code]
https://x.com/AlfieJCarter/status/2046985546279334105
A B2B GTM playbook that runs end-to-end inside Claude Code. Cowork does prospect research (project folder setup, web search, site screenshots, cold emails via Chrome extension without leaving Claude). Gmail + CRM connect in one session to draft 15 personalized follow-ups in under 10 minutes referencing actual call notes. Five routines run unattended: morning email brief, sales follow-up bot, LinkedIn scheduler, meeting prep assistant, weekly expense tracker. The DBS (Direction, Blueprints, Solutions) skill structure turns any repeatable workflow into a slash command.
@JJEnglert [Claude Code]
https://x.com/JJEnglert/status/2046961066437914681
A former LA filmmaker replaced After Effects and Premiere with a three-tool stack: Claude Code + HyperFrames + Descript. Motion graphics that took an hour per scene now take 3 to 5 minutes. Drop brand style guide into a folder, paste a transcript line into Claude Code, it reads the brand, generates a HyperFrames scene, renders to MP4, drags it onto the matching transcript line in Descript. Brand style guide is the unlock — Claude inherits fonts, colors, and reusable components across every scene without re-prompting.
@sumika45379 [Claude Code]
https://x.com/sumika45379/status/2046786103232078195
An Insforge Skills setup between Claude Code and its backend cut token consumption by roughly 3x. Before: 10.4M tokens, 10 errors, $9.21 per task. After: 3.7M tokens, 0 errors, $2.81. The trick is a context-design layer that preprocesses what Claude sees instead of dumping everything on it. The same pattern that collapses AI research budgets also collapses the "my Claude Code bill is out of control" anxiety thread under every Anthropic pricing post.
@defileo [Claude Code]
https://x.com/defileo/status/2047037849535762928
A $599 Mac Mini running Opus 4.7 on a single prompt — "What do you want the computer to do?" — repurposed as a dedicated always-on coworker. Fork a repo, write code, create a file, find an article, repeat. No laptop, no cloud subscription, no context switching. Someone looked at Apple's most powerful pocket computer of the last decade and turned it into a remote control for an agent that never clocks out. $599 once, never asks for a raise.
@garrytan [OpenClaw]
https://x.com/garrytan/status/2046981289031667961
Garry Tan says 50% of his agentic coding now runs on the /skillify pattern: do something once in OpenClaw, then tell it SKILLIFY IT, and the workflow becomes a reusable skill forever. This is the "my building process is itself my asset" insight made concrete. He's using it to build GBrain and a personal mini-AGI with full context on him and the things he cares about. The point isn't what he built — it's that the system remembers how he built it.
@mizchi [Claude Code]
https://x.com/mizchi/status/2046930090530075077
Someone asked how mizchi's PRs always auto-resolve conflicts overnight. The answer: Claude Code Web routine setup that runs Claude Code against every open PR in the middle of the night, resolves conflicts, pushes. Nobody is doing anything in the morning except looking at green checkmarks. This is agents-as-background-worker taken to its simplest form: one cron, one CLI, no infra, no LLM gateway, and the developer sleeps through merge conflicts for the first time in their career.
@coreyganim [Claude Code]
https://x.com/coreyganim/status/2046895911754752320
A client-ready website workflow without design or coding skills. Google Stitch generates design system + mockups from a business description in two minutes. Client approves. Stitch connects to Claude Code via MCP in 30 seconds. Tell Claude Code "Convert the landing page from Stitch" — it pulls the design system, downloads optimized images, builds a working Next.js site. Store the design system in design.md so every future page inherits the brand automatically. Full build under 30 minutes. Agency pricing still applies.
@learnwithella [Claude Code]
https://x.com/learnwithella/status/2046806088306270506
A Claude Code SEO agent replacing a $200/month Ahrefs subscription. Connects to Google Search Console, pulls real ranking data, finds "gap zone" keywords at positions 5-20, scrapes outranking competitors and analyzes why they win, interviews the user once about brand voice, writes content in that voice, tracks rankings weekly, feeds results back into the next cycle. Full loop. No expensive tools gathering dust, no freelancers producing AI slop that tanks after 90 days.
@learnwithella [Claude Code]
https://x.com/learnwithella/status/2046858357936685057
Same user, different experiment: self-improving Claude Code skills via an internal eval loop. Define 3-5 binary criteria. Claude runs the skill 10 times with varied inputs. A separate evaluator scores every output, identifies failure patterns, rewrites the skill prompt, retests. Winner stays. A hook-writer skill went from 32/50 to 47/50 overnight. This is Karpathy's autoresearch applied to marketing skills instead of ML training — same loop, applied to anything with a measurable output.
@aiwithjainam [Claude Code]
https://x.com/aiwithjainam/status/2046891782362042740
An open-source AI Legal Assistant repo that installs 14 skills and 5 parallel agents into Claude Code. It generates custom NDAs, GDPR/CCPA-compliant terms of service from a URL, auto-generated privacy policies, freelancer contracts, MSAs, and runs compliance audits across GDPR, CCPA, ADA, PCI-DSS, CAN-SPAM, SOC 2. Contract review with 5 parallel agents returns a Safety Score, negotiation plan, and PDF report in 60 seconds. LegalZoom charges hundreds per document; this does it in one install command.
@svpino [Claude Code]
https://x.com/svpino/status/2046928263134802314
The smallest useful pattern in this entire thread: `cat error.log | claude "fix this"`. Piping real context in beats copy-pasting a single error snippet and wondering why Claude is hallucinating solutions. The same trick drops straight into CI/CD — pipe failing job logs to Claude, get either an auto-fix or a detailed remediation plan. It's the Unix philosophy catching up with LLM ergonomics.
@rails [Claude Code]
https://x.com/rails/status/2047024401041019123
Intercom went all-in on Claude Code over one year: 95%+ of daily code is AI-generated, 1,000+ weekly internal users across sales, marketing, and finance (not just engineering), and engineering throughput doubled. Brian Scanlan shares the playbook on the On Rails podcast. The headline isn't the 95% number — it's that non-engineers are the biggest weekly user base. Claude Code isn't a dev tool anymore, it's an org-wide compiler for any written task.
@ai_xiaomu [Claude Code]
https://x.com/ai_xiaomu/status/2046800158797537373
A small team using Claude Code's Agent Swarm to run 16 parallel AI agents built a C compiler in Rust from scratch. 100,000 lines of code in 2 weeks. The pre-AI baseline for this kind of project was a small team for half a year minimum. The implication every engineering manager is trying to not say out loud: when one person plus Claude Code can run a 16-agent fleet on a systems-level project and finish in 2 weeks, the question isn't "how fast can my team go," it's "what is a team now."
@ai_hakase_ [Claude Code]
https://x.com/ai_hakase_/status/2047073976959013129
A local vLLM server paired with Claude Code via environment variables. MiniMax-M2.7 runs locally with Prefix Caching, and Claude Code talks to it at near-zero prompt-loading latency and zero API bills. Confidential code debugging and large refactors that would blow a Max plan weekly budget in a single session now run on the user's own machine. This is where the "Chinese open-weight models" conversation becomes practical: the scaffold is Claude Code; the brain is whatever you want to host.
@totoche [Claude Code]
https://x.com/totoche/status/2046938442974597579
15-year dev preparing the largest release of his SaaS taap.bio. Hasn't written a line of code in two years — Claude Code generates everything, he reads, comprehends, validates, occasionally rejects or retouches. Somewhere in those two years he became a product person without noticing or planning it. The 15 years of dev experience still matter; they just get deployed differently now. This is the least dramatic version of the "end of coding" debate and probably the most accurate one.
@VaibhavSisinty [Claude Code]
https://x.com/VaibhavSisinty/status/2046886165509476818
A Codex agent ran in full-auto mode for 9 hours overnight building a MacOS Content OS app. Full access to browser, filesystem, shell. 30-minute heartbeat automation reminded the agent it's supposed to stay autonomous, plus a VAPI phone-call tool if it got stuck. Hasn't made a single call. Figured out auth for YouTube Analytics and Meta Instagram Analytics by driving the browser for two hours. Integrated OpenRouter as an AI layer. Built a comment response engine using Apple Intelligence local models. The user has never written a line of code.
@doodlestein [Claude Code]
https://x.com/doodlestein/status/2047021115877281970
A bootstrapped one-person Claude Code skills marketplace (jeffreys-skills.md) hit $6k MRR six weeks after launch, growing ~200% month over month. No paid marketing, no venture funding, $20/month subscription (competing with $200/month Pro plans from OpenAI and Anthropic). All organic posts about new skills. The business model is: sell curated skill packs to the 4M+ Codex and Claude Code users who don't want to build their own. One person, one site, near-zero infra cost.
@BenjaminDEKR [Claude Code]
https://x.com/BenjaminDEKR/status/2047050265769934898
An ethical-hacking one-liner: WiFi camera's app is garbage, needs a remote "wake up" before streaming. Claude Code sniffs the BLE packets, finds the wake command, reverse-engineers it, writes a Raspberry Pi script to fire it from terminal. Total time: minutes. Ten years ago this was a weekend project for someone with hardware hacking experience. Now it's a conversation and a script.
@0xSero [Claude Code]
https://x.com/0xSero/status/2047083794750980313
A vibeproxy user wrote an observer that patches the Codex app at launch to inject models from Vibeproxy, so Claude can be used inside Codex and vice versa. The model-mixing layer is becoming its own product category. The interesting part isn't that it works — it's that Claude Code and Codex are converging into one substrate with different front-ends, and the most productive developers are already treating them as interchangeable with proxies.
@aakashgupta [Claude Code]
https://x.com/aakashgupta/status/2046953199345729881
A Claude Code PRD reviewer running Mahesh Yadav's actual PM checklist (urgency, differentiation from ChatGPT wrappers, AI failure modes, attribution risks) and dropping comments inside Notion. The compounding part: every 30 minutes a second agent reads Mahesh's edits to the AI's comments and writes them to learner.md. After five consecutive days of the same correction, it emails a proposed checklist update. He approves once. Every review gets permanently better. This is the loop PM roles are going to collapse into: review at the meta level, teach the agent the checklist, agent updates the checklist every night.
🗣 User Voice
User Voice
Users are furious but clear-eyed about what the Claude Code pricing drama actually means. @joelgascoigne called the A/B test "a business model change disguised as an experiment" that will win on short-term revenue metrics but erode brand trust for years. @jeremyphoward was blunter: "It's been a shock to see Anthropic's integrity collapse in the face of commercial pressure."
The deeper thread is compute scarcity honesty. @Yuchenj_UW put it simply: "Anthropic should just pay SpaceX $10B to buy GPUs. GPU scarcity is forcing them into all these weird moves." @soumitrashukla9 said the quiet part loud: "We are struggling to support Claude Code on the $20 plan because we simply don't have the compute."
Power users want scaffolding-first pricing, not usage-first. @bryancsk's viral line — "My Claude Code token budget at work has been increased and is now higher than my gross total compensation" — got the real follow-up from @KayvonJafar: "You are no longer the expensive part of the company. You are just the human attached to the inference bill."
Users are also tired of the scaffold-mismatch-as-regression story. @ChowdhuryNeil claims Opus 4.7 feels buggy because "nobody at Anthropic dogfooded it since they're all using Mythos." @m13v_ says most of the 4.7 regressions are "scaffold mismatch, not model quality — tool-call formatting and retry loops get tuned to the last model's quirks and silently break on the next one."
Users are furious but clear-eyed about what the Claude Code pricing drama actually means. @joelgascoigne called the A/B test "a business model change disguised as an experiment" that will win on short-term revenue metrics but erode brand trust for years. @jeremyphoward was blunter: "It's been a shock to see Anthropic's integrity collapse in the face of commercial pressure."
The deeper thread is compute scarcity honesty. @Yuchenj_UW put it simply: "Anthropic should just pay SpaceX $10B to buy GPUs. GPU scarcity is forcing them into all these weird moves." @soumitrashukla9 said the quiet part loud: "We are struggling to support Claude Code on the $20 plan because we simply don't have the compute."
Power users want scaffolding-first pricing, not usage-first. @bryancsk's viral line — "My Claude Code token budget at work has been increased and is now higher than my gross total compensation" — got the real follow-up from @KayvonJafar: "You are no longer the expensive part of the company. You are just the human attached to the inference bill."
Users are also tired of the scaffold-mismatch-as-regression story. @ChowdhuryNeil claims Opus 4.7 feels buggy because "nobody at Anthropic dogfooded it since they're all using Mythos." @m13v_ says most of the 4.7 regressions are "scaffold mismatch, not model quality — tool-call formatting and retry loops get tuned to the last model's quirks and silently break on the next one."
📡 Eco Products Radar
Eco Products Radar
Qwen 3.6 27B — Alibaba's dense coding flagship, open-sourced 2026-04-22. Targeted directly at OpenClaw and Claude Code as a drop-in model.
Kimi K2.6 — Moonshot's agentic coding model, tuned for long-horizon tool use; showing up as the Cursor Composer 2 foundation.
MiniMax M2.7 — open-source with "self-evolution" scaffold rewriting, showing up in local Claude Code via vLLM.
GLM-5.1 — ByteDance's ModelArk and BytePlus plans aggressively bundling it with Claude Code/OpenClaw/Cursor/TRAE for $10/month.
GPT Image 2 — OpenAI's new image model, ripping through Claude Code workflows for ad creative, slide decks, and UI assets.
/ultrareview — Claude Code's new cloud-parallel code review, free preview through 5/5 for Pro/Max, paid after.
Hermes Agent (Nous Research) — self-improving local agent, 111k GitHub stars, now running via `ollama launch hermes` and positioned as the OpenClaw successor.
HyperFrames (HeyGen) — HTML-to-MP4 video framework explicitly designed for agents instead of humans.
free-claude-code — open-source proxy that maps Claude API calls to NVIDIA NIM format for free model access at 40 req/min.
DESIGN.md / getdesign.md — Google's open design-token standard, being adopted by Claude Code and Cursor as a drop-in brand system.
abtop — GitHub terminal dashboard for multi-session Claude Code and Codex CLI monitoring (token usage, rate limits, context).
Monitor tool — new Claude Code primitive letting the agent sleep until an OS event wakes it up, killing polling-loop token waste.
/skillify — Claude Code's built-in "turn this workflow into a permanent skill" command, found inside the leaked source and rolling out.
Qwen 3.6 27B — Alibaba's dense coding flagship, open-sourced 2026-04-22. Targeted directly at OpenClaw and Claude Code as a drop-in model.
Kimi K2.6 — Moonshot's agentic coding model, tuned for long-horizon tool use; showing up as the Cursor Composer 2 foundation.
MiniMax M2.7 — open-source with "self-evolution" scaffold rewriting, showing up in local Claude Code via vLLM.
GLM-5.1 — ByteDance's ModelArk and BytePlus plans aggressively bundling it with Claude Code/OpenClaw/Cursor/TRAE for $10/month.
GPT Image 2 — OpenAI's new image model, ripping through Claude Code workflows for ad creative, slide decks, and UI assets.
/ultrareview — Claude Code's new cloud-parallel code review, free preview through 5/5 for Pro/Max, paid after.
Hermes Agent (Nous Research) — self-improving local agent, 111k GitHub stars, now running via `ollama launch hermes` and positioned as the OpenClaw successor.
HyperFrames (HeyGen) — HTML-to-MP4 video framework explicitly designed for agents instead of humans.
free-claude-code — open-source proxy that maps Claude API calls to NVIDIA NIM format for free model access at 40 req/min.
DESIGN.md / getdesign.md — Google's open design-token standard, being adopted by Claude Code and Cursor as a drop-in brand system.
abtop — GitHub terminal dashboard for multi-session Claude Code and Codex CLI monitoring (token usage, rate limits, context).
Monitor tool — new Claude Code primitive letting the agent sleep until an OS event wakes it up, killing polling-loop token waste.
/skillify — Claude Code's built-in "turn this workflow into a permanent skill" command, found inside the leaked source and rolling out.
Comments